Do ChatGPT, Perplexity, and Google AI Overviews Cite AI-Written Content Differently Than Human Content?

Short answer: yes, and the gap is bigger than most people expect. AI answer engines are not neutral about who wrote the source. They lean human, hard.

This is a different question from "does AI content rank on Google." That's a ranking question. This one is a citation question: when ChatGPT, Perplexity, or Google AI Overviews build an answer and pick sources to quote, do they treat an AI-written page the same as a human-written one? The research says no, and the data is remarkably consistent across studies.

1. The Headline Numbers

A study from Graphite analyzed content that actually shows up in Google Search and in AI answer engines. The results were direct: human-written articles made up 86% of pages ranking in Google Search, while AI-generated articles made up only 14%. It gets even sharper in AI answer engines specifically.

82% vs 18%

AI Answer Engine Citations

Of the articles cited by ChatGPT and Perplexity, 82% were human-written while only 18% were AI-generated.

86% vs 14%

Google Search Organic Ranks

Human-written articles made up 86% of top-ranking pages in Google Search, while AI-generated articles made up 14%.

That's not a small tilt. Human content is roughly winning citation share by more than 4 to 1, even though AI-generated content is now published online in higher volume than human content, according to the same research group's separate tracking study.

Key Takeaway

So it's not that AI content doesn't exist in large enough amounts to get picked. It exists in huge amounts. It's just underperforming on the metric that matters: getting quoted.

2. Why This Happens: It's Not Just Quality

Here's where it gets interesting. You'd assume this gap is purely about writing quality—AI content is worse, so it gets cited less. That's part of it, but new research points to something else going on underneath: actual bias toward the human label itself, not just the writing.

1. The Metadata Authorship Experiment (2025 arXiv Study)

A 2025 arXiv paper tested this directly. Researchers built a retrieval system where an AI model had to generate answers and cite its sources, and they controlled which documents were labeled as human-written versus AI-generated. Adding authorship information as metadata changed the model's citation and attribution behavior by 3% to 18%. The models showed a real bias toward citing sources explicitly marked as human-authored, separate from anything about the content quality itself.

2. The Blind Label Flip Test

A second study pushed this further and got a wild result. Researchers ran a blind test where evaluators, including AI models acting as judges, compared writing without knowing who wrote it:

Blind Unlabeled Test: AI content actually won a slight preference, chosen 55.3% of the time.
Correctly Labeled Test: The moment authorship labels were shown correctly, preference for AI content dropped to 47.8% (below the coin flip line).
Flipped Labels: When researchers mislabeled AI content as human, preference for that same AI content jumped to 61.5%!

"The wildest part: AI models judging other AI models showed this bias even more strongly than human evaluators did—more than double the effect size. So it's not just people being skeptical of AI writing. The AI systems doing the citing carry their own baked-in preference for the 'human' label too."

3. What This Means for Each Platform

Every platform doesn't work the same way, so the discrimination shows up differently depending on where you're trying to get cited:

ChatGPT

Leans heavily on consensus and established sources. Nearly half of its top citations point to Wikipedia, plus major outlets like Reuters and the AP. This structure naturally favors institutional, edited, clearly-human content over a fresh AI-written blog post with no track record.

Perplexity

Runs live web search on every query and pulls heavily from Reddit and recent content. It's less obsessed with big-name authority, but it still leans on human-generated community discussion as its top source category—which is basically the opposite of AI-written landing pages.

Google AI Overviews

Builds on the traditional Google index plus extra authority signals. Anything that already struggles to rank as AI-written content (per the earlier ranking data) starts one step behind before citation is even considered.

4. So What Should You Actually Do?

The move here isn't "never use AI to write." It's "don't let it stay obviously, purely AI." A few things the research actually supports:

Heavy human editing matters more than people think: The Graphite researchers specifically noted they didn't test AI-assisted content with real human editing, and they believe that combination could perform much better than raw AI output.
Original data and firsthand expertise: These are the clearest signals of human authorship an AI system can pick up on, since a generic AI draft has no real stats or lived experience to pull from.
Community and editorial trust still matter: Getting mentioned or discussed on Reddit, in the press, or through an established publication gives you the "human-associated" signal these systems are quietly rewarding.
Don't publish and forget: Freshness and continued human engagement (comments, updates, real discussion) reinforce the human-content signal over time.

5. The Bottom Line

AI answer engines do treat AI-written and human-written content differently, and it's not just a byproduct of quality gaps. Multiple studies now show a real, measurable bias toward human authorship baked into how these systems judge and cite sources, sometimes even stronger in AI judges than in actual people. If you're building content specifically to get cited by ChatGPT, Perplexity, or AI Overviews, the smartest bet right now is real human editing, real expertise, and content that earns human-associated trust signals, not just content that sounds human.

6. Sources

Graphite (Five Percent): AI Content In Search & LLMs Report
Graphite (Five Percent): More Articles Are Now Created by AI Than Humans Tracking Study
arXiv: Evaluation of Attribution Bias in Generator-Aware Retrieval-Augmented Large Language Models
arXiv: Everyone Prefers Human Writers, Including AI
Discovered Labs: ChatGPT, Claude, Perplexity, and Google AI Overviews: How Each Platform Cites Sources Differently
Leapd: How ChatGPT, Google AI Overviews, and Perplexity Source Information in 2026

7. Frequently Asked Questions

Do AI search engines actually treat AI-written content worse than human content?

Yes. Research from Graphite found that only 18% of articles cited by ChatGPT and Perplexity are AI-generated, compared to 82% that are human-written, even though AI content is published online in higher volume overall.

Is this discrimination just about content quality?

Not entirely. A 2025 study found that simply labeling a source as human-authored versus AI-generated changed an AI model's citation behavior by 3% to 18%, independent of the actual content quality, showing a real authorship bias.

Do AI models prefer human writing even when they can't tell who wrote it?

It's mixed. In blind tests with no authorship labels, one study found AI judges gave a slight edge (55.3%) to AI-written content. Once correct labels were shown, that preference flipped toward human content (47.8%), and mislabeling AI content as human made it most preferred (61.5%).

Which AI platform is hardest for AI-written content to get cited on?

ChatGPT appears toughest, since it leans heavily on Wikipedia and major established news outlets for its top citations, both strongly human-authored, institutional sources with long track records.

Does heavy human editing of AI-written content help with AI citations?

Likely yes, though it hasn't been fully isolated in research yet. The Graphite study specifically noted they didn't test AI-assisted content with real human editing, but researchers believe that combination could perform significantly better than raw AI output.

Does this mean I should stop using AI to help write content?

No. It means raw, unedited AI content is at a real disadvantage for AI citations specifically. Using AI as a drafting tool, then adding real human expertise, original data, and editing, is the strategy the research actually supports.

Is this authorship bias likely to change as AI content improves?

It's uncertain. Since part of the bias comes from the authorship label itself, not just writing quality, better AI writing alone may not close the gap. Building real human-associated trust signals matters just as much as writing quality.

Rishabh Kumar

Key Protocol Takeaways